Struct MultimodalConfig 

pub struct MultimodalConfig {
    pub max_images: usize,
    pub max_image_size_mb: usize,
    pub max_image_turns: usize,
    pub allow_remote_fetch: bool,
    pub vision_model_provider: Option<String>,
    pub vision_model: Option<String>,
}

Multimodal (image) handling configuration ([multimodal] section).

Privacy and cost note

Tool text is never scanned for image paths: a tool result that merely prints a local path (e.g. shell tools doing ls /pictures or find . -name '*.png') stays text. Image bytes are uploaded only when a tool explicitly declares an attachment, and a declared attachment is base64-inlined into the next provider request, so operators running tools that declare image attachments over personal or sensitive files should be aware of the upload semantics.

max_images (and the trim_old_images LRU policy) bounds the per-request image budget. See docs/book/src/contributing/privacy.md for the project’s privacy stance.

Fields

max_images: usize

Maximum number of image attachments accepted per request.

Caps the total number of [IMAGE:...] markers that survive into the provider request after multimodal preprocessing. Older images are dropped first. When a new image takes the count past the limit, the oldest images are removed in one batch, down to half the limit rounded up (4 -> 2, 8 -> 4, 16 -> 8); the count then grows back to the limit before the next batch. Removing an image changes the message that held it, which can invalidate a provider’s cached prefix from that message onward; batching means a session past the limit pays that once every max_images / 2 + 1 new images rather than on every one. Right after a batch the model sees fewer older images than the limit allows. A larger limit delays the first batch and makes each one less frequent. Acts as the upper bound on per-turn upload cost when tool outputs surface local image paths.

max_image_size_mb: usize

Maximum image payload size in MiB before base64 encoding.

Measured on decoded bytes, so the encoded request payload is about a third larger. Applies to every image entering the pipeline: channel attachments, tool outputs that surface local image paths, and the web dashboard upload. Defaults to the 20 MiB ceiling that MultimodalConfig::effective_limits clamps to, so an ordinary photo is accepted without configuration; lower it to bound per-turn upload cost or gateway buffering.

Providers apply their own limits on top of this one. Anthropic refuses a single image over 10 MB base64-encoded (about 7.5 MiB decoded), enforced by its provider client independently of this setting.

max_image_turns: usize

Maximum age of images in conversation turns.

When non-zero, images in user messages that are more than this many turns back from the end of history are stripped before the request is sent to the provider. This prevents a single screenshot from being re-encoded and re-uploaded on every subsequent turn indefinitely. Tool-result images are already managed by the stale-tool-result mechanism and are not affected by this setting.

0 (the default) disables age-based trimming entirely — images are only evicted by the max_images count cap.

allow_remote_fetch: bool

Allow fetching remote image URLs (http/https). Disabled by default.

vision_model_provider: Option<String>

ModelProvider name to use for vision/image messages (e.g. "ollama"). When set, messages containing [IMAGE:] markers are routed to this model_provider instead of the default text model_provider.

vision_model: Option<String>

Model to use when routing to the vision model_provider (e.g. "llava:7b"). Only used when vision_model_provider is set.

Implementations

impl MultimodalConfig

pub fn configurable_prefix() -> &'static str

Returns the #[prefix] value for this Configurable struct.

pub fn init_requires_explicit_config() -> bool

True when this struct has a required leaf field that a bare/ancestor-prefix init_defaults scaffold must not materialize (it would fill the field with its Rust Default, which prune_empty_leaves then strips on save, leaving a partial sub-table that fails strict reload). Explicitly targeting this section (or a descendant of it) still scaffolds regardless of this flag, so the section can be materialized for the operator to fill in.

pub fn secret_fields(&self) -> Vec<SecretFieldInfo>

Returns metadata about all #[secret] fields on this struct and nested children.

pub fn secret_field_terminals() -> Vec<&'static str>

pub fn encrypt_secrets(&mut self, store: &SecretStore) -> Result<()>

Encrypt all secret fields in place using the provided store.

pub fn decrypt_secrets(&mut self, store: &SecretStore) -> Result<()>

Decrypt all secret fields in place using the provided store.

pub fn set_secret(&mut self, name: &str, value: String) -> Result<()>

Set a secret field by its full dotted name, dispatching to nested children.

pub fn prop_fields(&self) -> Vec<PropFieldInfo>

Returns metadata about all property fields on this struct and nested children.

pub fn get_prop(&self, name: &str) -> Result<String>

Get a property value by its full dotted name, returning it as a display string.

pub fn set_prop(&mut self, name: &str, value_str: &str) -> Result<()>

Set a property value by its full dotted name, parsing from string.

pub fn prop_is_secret(name: &str) -> bool

Check if a property name refers to a secret field (static, no instance needed).

pub fn init_defaults(&mut self, prefix: Option<&str>) -> Vec<&'static str>

Instantiate None nested sections whose prefix matches. Returns the prefixes that were initialized.

pub fn map_key_sections() -> Vec<MapKeySection>

Enumerate every map-keyed (HashMap<String, T>) and list-shaped (Vec<T>) section discoverable from this Configurable’s tree. The dashboard / CLI consume this to surface “+ Add” affordances without hardcoding the section list.

pub fn nested_section_help(name: &str) -> Option<&'static str>

Help blurb for a #[nested] field on this struct, sourced from the field-level /// docstring. Returns None for unknown names so callers can fall through to a different lookup.

pub fn nested_section_group(name: &str) -> Option<&'static str>

pub fn get_map_keys(&self, section_path: &str) -> Option<Vec<String>>

Return the current alias keys at section_path, or None if the path doesn’t resolve to a map-keyed section in this tree.

pub fn nested_option_entries(&self) -> Vec<NestedOptionEntry>

pub fn create_map_key( &mut self, section_path: &str, map_key: &str, ) -> Result<bool, String>

pub fn delete_map_key( &mut self, section_path: &str, map_key: &str, ) -> Result<bool, String>

pub fn rename_map_key( &mut self, section_path: &str, map_key: &str, new_key: &str, ) -> Result<bool, String>

impl MultimodalConfig

pub fn effective_limits(&self) -> (usize, usize)

Clamp configured values to safe runtime bounds.

The 20 MiB image ceiling is the lowest common per-image or per-request bound across the supported vision APIs (OpenAI accepts about 20 MB per image, Gemini 20 MB for an inline request, Anthropic 32 MB per request with a tighter per-image limit its own client enforces). Images are buffered whole and grow by about a third under base64, so the ceiling also bounds gateway memory per upload.

Trait Implementations

impl Clone for MultimodalConfig

fn clone(&self) -> MultimodalConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for MultimodalConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for MultimodalConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for MultimodalConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl JsonSchema for MultimodalConfig

fn schema_name() -> Cow<'static, str>

The name of the generated JSON Schema.

fn schema_id() -> Cow<'static, str>

Returns a string that uniquely identifies the schema produced by this type.

fn json_schema(generator: &mut SchemaGenerator) -> Schema

Generates a JSON Schema for this type.

fn inline_schema() -> bool

Whether JSON Schemas generated for this type should be included directly in parent schemas, rather than being re-used where possible using the $ref keyword.

impl MaskSecrets for MultimodalConfig

fn mask_secrets(&mut self)

fn restore_secrets_from(&mut self, current: &Self)

impl Serialize for MultimodalConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.